In the meantime, here are some resources from the session which you could dive deeper:

• The slide deck, with the 5-Step AI Product Validation Framework:

  

• Roger’s 1:1 coaching and advisory: As an ICF-certified coach, I work with product leaders, founders and tech professionals build meaningful value with AI, and transform their career and life. Learn more at rogerjin.co

• Chris’s venture studio Arsenal AI is seeking early-stage founders and technology companies who want a venture partner that combines investment with hands-on execution, as well as enterprise clients ready to leverage AI for operational and financial impact. Learn more at arsenalai.com

I’d love to hear what you thought of the lesson. Email me at roger@rogerjin.co or DM me on LinkedIn with your comments and questions!

Cheers,
Roger

The key is validation – pressure-testing your idea early, rigorously, and with the right mindset.

From my own experience as a founder, product leader in Big Tech as well as coaching AI founders, I summarized my learnings into this 5-step guide. Each step is a mindset shift or critical decision point, to validate an AI product idea from scratch. Follow these steps to avoid common traps (like the infamous “solution-first” AI trap), craft a compelling AI-native value proposition, design for data and defensibility, iterate quickly with imperfect models, and leverage human-in-the-loop workflows to accelerate learning. Let’s dive in.

Step 1: Fall in Love with the Problem, Not the Solution (Avoid the AI Solution-First Trap)

The biggest sin in AI product development is starting with a “cool” AI solution and then wandering around for a problem to solve. Too many founders get overly excited with a new model or capability and forget to validate if anyone actually cares. In fact, lack of market need is the #1 startup killer – about 42% of startups fail because they built something no one wanted. AI startups are especially prone to this; it’s all too tempting to create an impressive demo that ultimately addresses no urgent user pain. For example, social robot Jibo garnered media hype but struggled to find a real market demand and shut down in 2018, a cautionary tale of an AI solution in search of a problem.

The mindset shift: Start with a specific user problem – ideally an acute, recurring pain point – before you even think about the letters “AI.” Identify a target user persona and a concrete job-to-be-done. Why is it painful or inefficient today? Why haven’t existing solutions fixed it? Ground your idea in that insight.

“Successful founders are in love with their problem, usually because they are in love with the customer... The ones that are in love with their idea and their product fail in massive numbers” - David Hirschfeld

In practice, this means writing a problem statement that resonates with a real customer’s frustration, without mentioning any AI magic. If you can’t clearly articulate who you’re helping and why they desperately need a better solution, stop and refocus.

Validate the pain early: Talk to potential users, even if informally. For instance, before building anything, you might interview 5–10 people in your target audience and ask about their current workaround or pain. Look for that telltale sign of a high-value problem: users expressing relief or excitement at the idea of a solution. Make sure you’re solving a significant problem (one that affects budgets, productivity, health, etc.) and not just a minor inconvenience. The goal is to avoid the trap of a shiny AI that’s a “nice-to-have.” In summary, start with an unmet need – the sharper and more specific, the better.

Practical checklist (Problem-First Validation): Before writing a single line of code, ensure you can answer these questions:

• Who exactly is the user, and what urgent problem do they face?

• How do they address it today, and why is that insufficient?

• If you offered a solution (even a manual one), would they realistically pay or invest time in it?

This problem-first discipline will guard you against building tech for tech’s sake. One AI founder recently scrapped months of coding after realizing he hadn’t confirmed if anyone would pay for his AI at all – he eventually went out and manually offered the service as a test, discovering that only then had he truly validated a real need.

Step 2: Craft an AI-Native Value Proposition (Focus on a 10× Advantage)

Once you’ve identified a real problem, the next step is framing why your solution needs to be AI-powered – and how that gives you a 10× advantage over the status quo.

In my previous post, I talked about the AI Trilemma Advantage, as a mental model to realize the super power of AI in service oriented solutions.

This is a mindset shift from thinking “we’ll sprinkle in AI to sound innovative” to designing an AI-native value proposition from the ground up. Ask yourself: In what way can AI solve this problem fundamentally better or differently than traditional methods? The best AI product ideas aren’t just a little more efficient; they redefine the solution space by breaking old trade-offs. AI can often do things that were previously impossible – like providing personalization at scale, understanding unstructured inputs (language, images) automatically, or continuously learning to improve outcomes. Your value prop should hinge on one of these unique capabilities.

Think of it this way: if removing the AI from your product still leaves a viable solution, you probably haven’t pushed far enough. An AI-native product should deliver something only AI can do, or at least be order-of-magnitude better in speed, cost, or quality. For example, GitHub Copilot gives developers an “AI pair programmer” that can autocomplete code and suggest functions as they type. Just two years after launch, Copilot is already writing almost 50% of the code in projects where it’s enabled – a staggering leap in developer productivity that simply would not be possible with a non-AI tool. Another example: Khan Academy’s new tutor bot, Khanmigo, offers personalized, conversational tutoring to every student. Scaling one-on-one human tutoring to millions was impossible before; with AI, Khanmigo can coach students individually, at any hour, for virtually zero marginal cost. That’s an AI-native value proposition: something that was previously unattainable (a tutor for every child) is now within reach, courtesy of AI.

Articulate the “AI superpower”: To craft your AI-native value prop, clearly define what competitive edge the AI provides. Is it hyper-personalization (e.g. an app that adapts uniquely to each user’s behavior in real time)? Is it scalable insight (e.g. analyzing thousands of data points or documents in seconds to give an answer)? Is it creative generation (producing content, designs, or code on the fly)? Make that the centerpiece of your pitch. For instance, instead of saying “Our product uses machine learning to improve marketing emails,” you’d say, “Our AI copywriter drafts tailored emails for each customer segment in seconds, something a human team would take weeks to do.” The difference is framing the user benefit that’s unlocked by AI. A strong test is the 10× rule: if your AI solution isn’t at least 10 times faster or 10 times cheaper or enabling something 10 times more effective, go back to the drawing board. AI for AI’s sake won’t sell – users flock to AI products that feel like magic, not those that feel like ordinary products with a bit of predictive text thrown in.

Finally, avoid the trap of incrementalism. Being “AI-powered” is not a selling point on its own in 2025 – it’s about what new value AI delivers. So zoom out and describe your product’s value proposition in one compelling sentence. For example: “With our AI-driven recruiting tool, a hiring manager can screen 1,000 resumes in an afternoon and pinpoint the top 5 candidates – a process that used to take weeks.” That highlights a step-change in capability. Craft a similar AI-native value prop for your idea, and you’ll have a clear beacon to guide product development and messaging.

Step 3: Design for Data and Defensibility from Day One

In AI products, data is your moat – if you design it right. Unlike classic software, an AI product doesn’t just ship code; it continuously learns from data. This means two things: (1) Your product should be engineered from the outset to capture the critical data that makes its AI smarter, and (2) The way you collect and leverage data will largely determine your long-term defensibility against competitors. A common misconception is that you need a huge proprietary dataset upfront to have any chance at success. In reality, the new competitive moat is not about hoarding a static trove of data – it’s about establishing a dynamic learning loop. The most valuable data is the stream of interactions and feedback from your own users, which competitors can’t easily copy. One expert insight put it this way: a startup with just 1,000 engaged users feeding a system with high-quality feedback can build a stronger moat than a giant firm with a billion generic data points. It’s not the size of your dataset; it’s how fast and smart you are at learning from it.

Plan a data flywheel: Think through the workflow of your product and identify where you can capture inputs, outputs, and feedback. For example, if your AI generates a recommendation or prediction, will you let users rate its accuracy or correct it? If so, those user corrections are gold – they can flow back into model training or evaluation. Designing for data might mean adding a thumbs-up/down button on an AI-generated answer, a quick survey after an AI-driven session, or instrumentation that tracks success/failure of AI actions. Every interaction should, ideally, produce a datapoint that helps improve the model or the overall experience. This is “designing for the feedback loop.” Companies like Google mastered this: every search query and click refines their search algorithms. Waze, the navigation app, famously turned its users into sensors – each drive with the app on contributed traffic data to improve route suggestions for everyone. These are data network effects in action: the more people use the product, the better it gets. From day one, ask how your AI idea can leverage a similar effect on a smaller scale. For instance, can your AI product learn a little bit from each task it does for a user, so that tomorrow it performs even better? If yes, outline that mechanism clearly.

Build defensibility early: Why is this so critical? Because core AI algorithms and even large pre-trained models are increasingly commodities – what’s to stop a competitor from taking an off-the-shelf model and replicating your features? Your defense is the proprietary data and insight you gather over time. Consider the fate of Lensa AI’s avatar generator: it went viral using an open-source Stable Diffusion model to create artistic portraits, but within weeks a swarm of copycats (Dawn AI, Wonder AI, etc.) appeared with similar capabilities. Lensa had no lasting moat because the underlying tech was not unique and it wasn’t building a novel data loop beyond the model itself. To avoid “day 0 commoditization,” design a moat that strengthens as you grow: for example, a unique dataset (like labeled medical images with expert annotations that only you have), a community or network that produces exclusive data (user interactions or content locked into your platform), or a continuously improving model refinement process (like fine-tuning on user-specific data).

Be mindful of data quality as well. Often, quality beats quantity when it comes to training data. A small, well-curated dataset can outperform a massive noisy one in driving better model performance. As AI pioneer Andrew Ng emphasizes, systematically improving your data (fixing labeling errors, ensuring consistent definitions) can turbocharge an AI system without needing more data points. So early on, identify the critical data that your AI absolutely needs to get right. For a computer vision product, it might be collecting images of edge-case scenarios. For a language AI, it might be gathering example queries or dialogues from real users in your domain. Focus on those and instrument your product to capture them.

In summary, treat data as a first-class design concern. In your product roadmap, include features whose main purpose is to generate or enhance data. It could be as simple as an onboarding flow that asks new users a few key questions (feeding your model’s understanding), or as involved as a “labs” feature where power-users correct the AI’s mistakes and thus label new training examples. The payoff is twofold: you improve your AI rapidly, and you create a defensible asset that grows over time. As one venture capitalist noted, data network effects mean that over time nobody can serve your users as well as you can, because your product has literally learned from millions of interactions that competitors never saw. That’s the ultimate goal – a self-reinforcing cycle where more users → more data → better AI → more value → more users, and so on.

Step 4: Iterate with Imperfect Models and Fast Feedback Loops

Traditional product development might spend months polishing features before exposing them to real customers. With AI products, that approach is a recipe for wasted effort. AI systems are probabilistic and complex – you won’t know exactly how your model performs in the wild until you get it in front of users. So, embrace a new mantra: launch early, launch often. Your initial AI model will likely be rough around the edges, and that’s okay. In fact, it’s expected. Rather than aiming for 99% accuracy out of the gate, decide what “good enough” looks like to start learning from real usage. Often, 80% accuracy is a sensible initial target for an AI MVP (Minimum Viable Product) – it’s sufficient to deliver some user value and get feedback, without chasing diminishing returns. As one product leader puts it, the goal of an AI MVP is validated learning, not perfection.

Why is launching an imperfect model not just acceptable but desirable? Firstly, users can often tolerate imperfections if the overall value proposition is strong. A great example is Khan Academy’s AI tutor, Khanmigo. Students testing it have said, “I love Khanmigo. Yeah, every now and then it makes an error, but I don’t know what I would do without it.”. In other words, if your AI saves users significant time or effort, they’ll forgive the occasional glitch – especially if you’re transparent and responsive about improving. Secondly, early user interactions with your imperfect AI will highlight exactly where it falls short. Those insights are priceless; they tell you what to prioritize. Maybe the model’s accuracy is fine for 90% of cases but fails on a crucial 10% – now you know where to focus additional training or whether to add a rule-based fix or a UI tweak. You simply cannot get this feedback if you delay real-world testing. And thirdly, getting a prototype in users’ hands quickly helps manage expectations and training. Users often need to learn how to use a new AI tool effectively, and their behavior will adapt over time. By involving them early, you’re effectively co-evolving the product with its users.

Set up fast feedback loops: This is where an evaluation-driven mindset replaces a feature-driven one. Instead of adding a bunch of features up front, you build the smallest possible product that can test your core hypothesis. For example, suppose your idea is an AI that summarizes legal documents to save lawyers time. A fast-loop approach would be: build a bare-bones interface where a user can upload a document and get an AI-generated summary – nothing more. It might use a generic model (like GPT via an API) with a few dozen fine-tuning examples. Give it to a handful of friendly users and measure everything: Does the summary capture the key points? How often do users have to edit it? How long does it take them to review it versus reading the full document? This closed-loop feedback (user outcome data + qualitative feedback) is your guiding light. Maybe you discover that the summaries are accurate on simple contracts but falter on complex ones – that’s a signal to refine your training data for those complex cases. Or perhaps users want a way to highlight sections that must be included in the summary – that could inform a feature update. The point is, each iteration should cycle quickly: hypothesis → prototype → test → learn → refine. Many AI teams aim for weekly (or even faster) iteration cycles for this reason.

A helpful practice is to instrument your AI service with evaluation hooks. For instance, if you have an AI output, allow users to rate it or mark it as correct/incorrect. Track objective metrics like accuracy, precision/recall, or user task success rate on each new version of the model. Make the feedback loop as tight as possible. If feasible, do a staged rollout: try the update with 5 users, learn, then 50 users, and so on. You might maintain a sandbox or beta program where power users see new AI improvements first and give rapid feedback. Remember, with AI, data is the new debugging. Instead of stepping through code, you’re examining where the model’s predictions went wrong and why. The faster you get that data, the faster you can fix or improve it.

One more tip: don’t shy away from communicating the “beta” nature of your AI to users in early stages. Many will appreciate that they are part of shaping a cutting-edge product. For example, when Gmail first introduced their AI “Smart Compose” feature, it wasn’t perfect, but users understood it was learning from their usage. Framing your product as an evolving system can actually build user loyalty – people love being early adopters contributing to progress. Just ensure you close the loop by acting on feedback and showing improvements. Users will trust your AI more when they see it getting better over time in response to their needs.

In summary, speed trumps polish in AI validation. Get a working slice of your product out quickly, even if it’s only semi-automated or uses a lightweight model. The real-world lessons you gain will far outweigh the discomfort of not being perfect. And ultimately, those rapid iterations will converge your product toward something that truly nails the user’s problem – which is far more important than an AI that’s academically impressive but practically irrelevant.

Step 5: Leverage Human-in-the-Loop (HITL) as Your Secret Validation Weapon

“Human-in-the-loop” isn’t just a safety net for AI quality – it’s a cheat code for faster validation and development. The idea is simple: use humans to augment or monitor the AI system, especially in the early stages, to ensure your product delivers value and to accelerate your learning. There are two primary ways to leverage HITL when validating an AI idea: (a) behind-the-scenes humans who perform parts of the task that AI can’t yet do well (often called a “Wizard of Oz” prototype when the user is unaware), and (b) human oversight of the AI outputs, where people review, correct, or approve the AI’s work before the user sees it. Both approaches can dramatically shorten the time to get a working solution in users’ hands.

Why is this so powerful? Think back to our Step 1 emphasis on validating the problem and solution. If you can deliver the core experience of your product with some manual work behind the scenes, you should absolutely do so rather than wait to perfect the automation. This is exactly what savvy AI founders do. Consider a real example: Google Duplex, the AI system that calls restaurants to make reservations. In demos it wowed everyone by sounding human, but in practice Google didn’t rely on AI alone. From the start, they paired Duplex with a “human fallback” team. If the AI got confused or the conversation went off script, a human operator would silently take over the call. Those human operators also annotated the call transcripts to feed back into training data for Duplex. The result? Users got their reservations made seamlessly (high service quality from day one), and Google rapidly learned from every failure case to improve the AI. Duplex’s rollout was carefully managed with humans in the loop until the model could handle about 80% of calls end-to-end on its own. This blueprint can apply even on a smaller scale: whatever your AI can’t do yet, see if a human can step in so the user experience is complete. You’ll gather invaluable data on what real usage looks like, and you won’t lose users to early AI hiccups.

For early-stage startups or product teams, a Wizard of Oz prototype is often the quickest way to validate the whole product concept. For instance, if you’re building an AI medical advice chatbot, you might initially have a human doctor or medical student sitting behind the chat interface, crafting responses (or at least vetting AI-generated responses) without the user knowing. This lets you test: Do users actually find value in a 24/7 chat where they can ask health questions? What do they ask most, and what answers satisfy them? You can simulate the AI’s presence long before the AI is fully ready. Crucially, this isn’t “cheating” – it’s doing the scrappy manual work to prove (or disprove) that your idea works. As an added benefit, the transcripts from these interactions become training data for the eventual model. There’s a famous mantra in startups: “Do things that don’t scale” at the beginning. In AI, we modify that to “Do things manually before you scale with AI.” If 100 users love your service when you’re secretly powering it with humans, you’ve hit on something. Then you earn the right to figure out automation. As we saw earlier, one founder offered a manual checklist service for home kitchen inspections (charging a fee and doing the work himself) before building any AI – he validated that people would pay for the solution without a single AI model in place. Only after that proof did he start automating parts of it.

Design HITL workflows into the product: Beyond prototypes, think about your live product’s launch version having a human-in-loop component. This could mean, for example, moderation – if your AI generates content, have humans review the outputs initially to ensure they’re correct and appropriate before releasing to users. It could mean on-demand human assistance – if your AI is unsure or below a confidence threshold, route the task to a human expert and deliver that result to the user. Many “AI” services are actually AI-human hybrids under the hood, especially in sectors like legal, healthcare, or customer service where accuracy is paramount. Users don’t mind, as long as their problem is solved. In fact, highlighting that experts are supervising the AI can increase user trust. Importantly, these humans aren’t just there to put out fires – they are part of your learning loop. Each time the AI falters and a human corrects it, that’s a lesson for your model. Structure how you capture those lessons (e.g. logging the AI’s output, the human’s correction, and feeding that into your retraining pipeline).

Another HITL tactic is crowdsourcing for edge cases. If your AI needs a lot of labeled data, you can integrate crowdsourced labeling into the validation process. For instance, if you have an AI that classifies expense receipt images, you might deploy it at 50% capacity and have crowdworkers (via a service like Mechanical Turk or Scale AI) verify or label the rest in real time. This way, your early users always get a result (half from AI, half from humans behind the scenes), and meanwhile you’re rapidly building a labeled dataset of receipts to train a better model. This approach can be budget-sensitive, but it can bootstrap an AI in areas where data is scarce.

The overarching mindset here is don’t let AI’s limitations block you from shipping. Use people as the bridge to cover those gaps initially. It requires humility – your “high-tech” product might rely on some very low-tech processes at first – but it’s incredibly effective. You maintain user momentum, validate the end-to-end experience, and usually can do so at a fraction of the cost and time it would take to build a fully automated system that might miss the mark. Just be conscious to transition or scale the human elements thoughtfully: as your model improves, you can gradually reduce human involvement or shift them to handle only the most complex cases. In the long run, you might even keep some human touch for premium service or oversight (think AI medical diagnostics that always get a human doctor’s second look for critical cases).

To recap, human-in-the-loop is your ally, not your enemy. It can accelerate your path to product-market fit. By combining the creativity and empathy of humans with the speed of AI, you get the best of both worlds in the early days. So ask yourself: what’s the simplest version of my product with a human in the loop that I can test right now? Do that, and you’ll learn far more in a month than many teams learn in a year of building in isolation.

Conclusion: From Validation to Scale

Validating an AI product idea is as much about mindset as it is about tactics. By adopting a problem-first, solution-second mentality, you ensure you’re working on something meaningful. By demanding an AI-native value prop, you push your concept toward real differentiation and impact. By designing for data and defensibility from the start, you set up a flywheel that powers lasting competitive advantage. By shipping early and iterating with imperfect models, you tune into reality and outlearn the competition. And by putting humans in the loop, you combine the best of human insight and machine efficiency to accelerate your validation exponentially.

For product leaders and founders in the AI era, these steps are not a one-time checklist but a repeatable cycle. As you go from idea to MVP to scaling up, you’ll revisit these principles – avoiding new solution-first temptations, evolving your value proposition as technology advances, continuously investing in your data moat, tightening feedback loops, and recalibrating the human/AI balance in your system. AI-native product development is a journey of continuous learning, much like the models we train. The five steps above will help you navigate the critical early decisions so you build the right product in the right way.

In the end, remember that successful AI products emerge from a blend of strategic depth and practical execution. Keep your eyes on the strategic prize (a product that truly changes the game for your users) while ruthlessly staying practical (test assumptions, measure, adjust). If you do that, you won’t just validate your AI idea – you’ll set the foundation to launch and lead in this exciting new frontier of AI-native products. Good luck, and happy validating!

We live in a society where many of the historically stigmatized topics are now less and less taboo, such as sexuality, gender identity, mental health, etc. The drastic evolution of our beliefs and values in just over a generation of time has changed our culture significantly, even though there’s still plenty of stigma and biased opinions. In social or public settings, people can talk freely about politics, religion, spirituality, sexual orientation or health issues (like the use of Ozempic). However, very rarely do people discuss money issues, like their income or wealth. Money is still one of the last frontiers we keep close to ourselves.

It is because we don’t just hold money in our hands. We let it get into our head, and too quickly it gets to our heart.

In my work as a coach, money frequently comes up as a significant factor affecting people’s professional and personal decisions, but often in an unconscious way. People usually ignore, dismiss, or suppress the thought of money issues, hesitant to approach it directly. In this blog, I’m dissecting this topic into three aspects: economic, self-worth, and emotional, establishing a mental model that could help us uncover our unconsciousness. When we feel more comfortable calling out the elephant in the room, we are not to be blindsided by it and can make decisions with a clear mind.

Money in your hand

Money in your hand is for economic means, its basic functionality. “I have to pay the bills.” Many of my clients live in the SF Bay Area and New York. The living cost has been going up and seems to continue in the foreseeable future. To make ends meet takes a lot of hard work for many families. It is undeniable that financial pressure is a real thing for people; however, the majority of the sufferings that money costs us are actually not its pure financial implications. Let’s take a look at what they actually entail.

Money in your head

When you extend money beyond its economic means, you step onto a slippery slope. It gets from your hand to your head. Your mind starts to over-identifying with money. Before you realize it, you unconsciously equate money to your self-worth. Your annual total compensation package needs to go up because how can you be worth less as you work so hard. You are worth no more or less than your money. Then yes, of course the more the merrier. As you could imagine, this proxying could be very unhealthy. As life happens, any fluctuation in your money shakes your self-worth. Tying your identity to money is like putting fetters on yourself. You surrender your freedom and allow external forces to drag you anywhere they are going. Guess what? They are not going to always bring you to pleasant places.

Money in your heart

People not only hold money in their hands, think about it in their head, they also take it very deep into their heart. They need it as a drug to stabilize their emotions. Possessing or earning more money might be necessary to make you feel safer; while making less could be an emotional trigger for fear, anxiety and panic. These emotions do not come from simple financial calculations; instead they arise from very deep in the psyche. It’s very much like dependence on a drug. You need to continue taking it to feel centered; and not taking it could immediately invoke withdrawal symptoms. This unconscious perception of money is often rooted in a person’s family and cultural background. Growing up from a lower socioeconomic status with a constant lack of money could have a lasting impact on how people perceive money, even if later in life they become much more financially stable. This happens quite a lot for first-generation immigrants or self-made high-achievers who had to pursue their dream starting from scratch. They might have “made it” or realized their dream from an objective point of view; however, they continue living and operating in a dream mode, unconscious about how their emotional dependence on money still drives their every decision and action.

How to maintain a healthy relationship with money

As we discussed various ways that people could become unconscious on how they perceive money, how shall we cultivate a more conscious view and maintain a healthy relationship with money? I invite you to a very simple practice. Next time when you have an uncomfortable feeling about money, such as making a large purchase, filing your tax return, or asking for a raise from your employers. Notice your thoughts and emotions such as doubt, excitement, fear, dreadfulness, or whatever it may be. Take a moment to recognize it and check the underlying motivations. There might not be right or wrong thoughts and emotions. The key is to bring awareness of your true perception and belief, which is hugely beneficial to help you make a sound decision while cultivating a healthy and balanced relationship with money.

Keep money in hand. Be very careful when it gets into your head and your heart.

Ever since I was very little, I was a very goal-oriented person. I set the target and I achieved it. I created my bucket list and I checked it off. Happiness used to pop up very often whenever a goal was reached, and a box checked. The frequency of these little successes, with the cultural and societal affirmation and years of reinforcement, set up a fixed mindset for me that happiness equals to achieving your goals.

When I was 25, I created a great bucket list that I was very proud of: marrying my then-girlfriend, becoming a father, starting my own company, and making my first million dollars. I knew, with 120% certainty, that ever-lasting happiness will be mine to have if I achieve these goals.

When I turned 35, I checked all my bucket list. I have a loving wife and two kids. I was the founder of two successful startups. I made more money than I dreamed.

However, I was more unhappy than ever.

My anxiety and depression prompted me to go on a relentless search. I searched externally and internally. If I could find something wrong, I’ll be sure to fix it. If I could find something missing, I’ll be sure to get it.

My search went on for years, but to no avail.

Then it suddenly dawned on me, that there‘s nothing wrong, and there’s nothing missing. It is all because I have accepted from society a wrong definition of happiness. I’ve been chasing the wrong thing.

Our society defines happiness as the experience when you get to your destination. You feel happy when you score straight As at school, accepted into a prestigious college, get a promotion, a raise, a lot of money, etc. The hero beats the villain and marries the beauty, then they are happy, “ever after”.

Happiness is a feeling of “high”, a fleeting experience. By definition, it doesn’t last. You get it and you lose it. In fact, according to how our biology works, the high tanks to a low sooner or later, and then you want more. If you are into that feeling, you are never going to be content.

Are we supposed to never be content? Our society is sustained by discontentment, which pushes our consumerism economy forward. It doesn’t value contentment. The reality is that, as a culture, we are against contentment. Collectively, we are afraid that if we are content, we ain’t going to grow. If we have lasting satisfaction, we are never going to improve.

Are we able to break that cycle? Can we get off this happiness hamster wheel? Are we going to avoid becoming lazy and unproductive? Yes, yes and yes. Actually we would become the best version of ourselves.

I’m not saying that we don’t pursue happiness. Instead, a redefined happiness. Instead of fleeting excitement, we aim for lasting peace and joy. Instead of a singular outcome, we strike a holistic balance in approaching our work and life. Instead of obsessing with a tunnel-visioned goal, we establish a harmonic relationship with ourselves and the world.

Happiness is not a destination.

It is a journey.

And I invite you to join this journey.

Founders in the past are like Columbus in the 1400s.

They start the journey with the state-of-the-art technologies with a vision of what that voyage could take them within the next few years.

Founders in today’s AI world are a different version of Columbus: they set sail with wooden ships, but knowing that in a few months the fleet would upgrade to steam engines, and later into gas-powered cruise ships. In the last leg of the venture, they would be captaining a nuclear carrier with a crew of jet planes surveying the destinations.

That’s what today’s founders would look like.

State-of-the-art technology is what you start with; but expect that you’d soon be propelled by forces you didn’t have a year ago. At the same time, you’d surely need to swap the engine while you fly. The execution required is on a whole new level, so is the outcome.

Founders, imagine you’re Columbus with that vision.

“I’ve already thought about quitting my job three times today! And it’s only lunch!”

I received this text from Brian, a client of mine who is a product management lead in a tech company based in SF Bay Area.

When I saw Brian in our session two days later, not to my surprise, the floodgate opened immediately.

“The leadership team has been treating me unfairly! And so-called promotion path was bullsh*t. The culture in the team sucks and I’m not learning anything new in these stupid projects. But I’ve put in so much into this job that I feel so stuck!”

After he cooled down a bit, we explored deeply what these complaints really entail, the meaning of his career, and what learning and growth would look like.

“If you are not able to have all the positive things in a job, what is most important to you?”

With enough reflection, it happened that what Brian really looks for is the opportunity to lead consumer-facing AI products. Team size and promotion are nice to have, but not blockers to advance his career. Leaving his current position would still be a big loss, but now he knows exactly what he’s pursuing by giving it up.

Empowered by that clarity and resolve, Brian proceeded with switching to a new AI team within the same company, managing a smaller team but diving into a more exciting product area. It turned out to be a great move and he almost became a new person in his new job.

Don’t simply look for a new job because you’re stuck.

Get unstuck first, and then the best job will come to you.

Note: Explicit consent was given from the client to share this story, and the name and related details were altered to respect confidentiality.

In the previous post, I talked about Achievement Addictive Disorder, using myself as a case study. Granted, working hard is surely a merit to be encouraged and celebrated in our society. It is when things go to the extreme that causes pathological problems. Some of the AAD symptoms can be found in a lot of the high achievers, entrepreneurs and professionals working in fast-moving industries like tech and finance. Cultural, social and environmental influences on the side, there are psychological and emotional aspects that contributed to the individual’s tendency to suffer from this condition. In this post, I’m sharing a few prevalent issues that I found highly relevant to achievement addictions.

Note: The names and details of the coaching examples below are composites and have been altered to respect confidentiality.

Lack of Self-Love and Acceptance

People seek love and acceptance from others, but we also need constant, non-judgemental self-love and self-acceptance. I can’t be more familiar with this yearning. I grew up in a familial and cultural background that I have not learned to maintain a level of healthy self-love. When I do not find enough love and acceptance from myself, which is very often, I try to boost it using external factors. Work is a perfect endeavor to cultivate self esteem. It’s absolutely fine until I put in so much time and effort, and associate my entire identity with work. My professional achievement equates to my self worth. The worst part of it is: the more you work, the more you hold on to this illusion.

Marie, a coaching client, is a successful marketing professional and has a loving family, but frequently finds herself sabotaging her relationship by overworking.

Marie: I work at least 70 hours a week and sometimes 80 at my job. I don’t have any energy left to spend with my family and kids. My husband has been complaining all the time and I haven’t had any friends for a long time.

Me: You surely put all you can into work; what does that mean to you?Marrie: I cannot not work that much. If I don’t, I feel worthless. What have I achieved this week? If I don’t get these things done, why am I even here?

Marie: I cannot not work that much. If I don’t, I feel worthless. What have I achieved this week? If I don’t get these things done, why am I even here?

Social Judgement and Comparison

As social animals, we look for love, approval and appreciation from others all the time. However, without enough consciousness, our seek turns obsessive. Over the years of my tech career, perfectionism and fear of failure have been my two worst friends. Overworking with the goal of doing things perfectly, to be better than others or impress others, actually reduces the efficiency, and sometimes the quality of work itself. Fear of failure acts as a constant, annoying reminder on the side every time you try to loosen the grip of this addiction. Fear could even distort your understanding of reality. Unconsciously, I put my own negative, judgemental self-talk into the mouths of people around me. Instead of seeing others as collaborators and supporters, they come up as commentators, and worse yet, judges.

Social comparison could be another driver for people’s AAD. The root of this psychological tendency can be found in early years of one’s life, adolescent or even younger, when peer pressure from friends at school might have caused troubling memories.

Patricia had a successful career in her current company, being promoted to Director of Business Development last year.

Patricia: I’ve been working non-stop for the last six months, day and night. But I still feel that’s not enough.

Me: What makes you believe that you are not doing good enough.

Patricia: Lauren, the new VP, is the same age as I am. I fear that I’ve not been on a faster trajectory and I’ll regret it in a few years.

Escape or Avoidance Mechanism

Just like alcohol and other addictive substances, overworking is really “effective” in helping people escape from their life’s challenges and troubles. When immersed in work, enjoying the highs and indulging the lows, you don’t have to face those “hard” problems in life, like your relationships, family issues, and sometimes your health. You put on your professional mask and lose yourself in it. You wear this VR goggle and let the 360-degree experience take over your mind. The bad news is, just like any substance, you have to take off this mask and goggle, and experience the unpleasant withdrawal. Worse yet, your other problems in life are not solved.

Giovanni is trying to save a struggling relationship with his wife, and is also frustrated that both his teenage kids are estranged from him.

Giovanni: I’m very excited about this new opportunity to move to the global business unit. I will be able to expand my scope. I’ll have to split my time between London and California, and manage an important part of our European business.

Me: That definitely sounds like a great opportunity career wise. What would it mean to you and your family?

Giovanni: I haven’t thought about it yet. Do you think I should discuss it with my wife?

There are of course other reasons why hard working ethics turn into addictive and pathological. Some people are simply afraid of making a change, either adjusting the way they work or the environment they work in. A familiar feeling of being stuck appears to be more acceptable than embracing any unknown. Another common reason is financial insecurity, in a pure economic sense or emotional perspective, which I will explore further in my next blog. Stay tuned.

I rarely drink, and have never done drugs at all. I wouldn’t consider myself having been an addict in my life. Until recently.

In my master of counseling program, one of the core courses was abnormal psychology and psychopathology, in which we learned about diagnostics on various substance abuse and addiction related disorders. I was surprised to find striking similarities between these addictive pathologies and how I approach my professional life. As I reflect more on the way I perceive and pursue achievements at work, I came to realize that it has not only been unhealthy but also in many ways debilitating, much like substance-related addictions.

As defined in the Diagnostic Statistical Manual of Mental Disorders Fifth Edition Text Revision (DSM-5-TR, which is considered as the bible of mental health), substance related disorders encompass 10 separate classes of drug, including depressants such as alcohol, stimulants such as cocaine, and other drugs like opioids. If I swap the substances with how I chase professional accomplishments with excessive workaholism, it becomes clear that I may have what I call Achievement Addictive Disorder (AAD). As a person who devotes almost everything in pursuing so-called success, I have put other areas of my life in jeopardy and only recently began to realize the pathological nature of this behavior. Below I present myself as a case study of AAD, with the intention to raise awareness of an issue that is too often overlooked in our work culture. An important note here is that this disorder, like many other mental illnesses, is not necessarily binary (either you have it or not) but rather dimensional: everyone is on a spectrum with variance in degrees.

The diagnosis follows the same criteria on substance related disorder in DSM, which can be bucketed into three categories:

I. Large Amount

• Using larger amounts or for longer time than intended

Starting from my graduate school at Carnegie Mellon and throughout my career in the tech industry as an engineer, entrepreneur and product leader, I have been constantly working for long hours. If you ask me how many hours I work every week; the truth is that I don’t count it. For most of the last ten years, I was working whenever I could find time.

• Persistent desire or unsuccessful attempts to cut down or control use

This is an important criteria for any addictive disorder, in that the person has an intention to change their habit but fails to do so. Yes, I am enthusiastic about the work itself. It was intellectually challenging and I have a strong sense of community contribution. However, when the hour becomes grueling, enthusiasm fades and the unconscious momentum keeps the wheel rolling. Work began to take its toll on me, and I have had multiple times that my physical, emotional and relational life got into big problems; however, I still was not able to cut back on the time and energy I put into work.

• Great deal of time obtaining, using(in this case, working), or recovering

People with substance related disorders find their lives circling around the substance, from obtaining, using and recovering. I found myself revolving around professional achievement. If I’m not overworking, I’ll be recovering from the suffering of overwork. In some rare intermissions, I’m busy setting up the next ambitious goal to overwork towards.

Up until this point in the list of symptoms, it could still seem “normal” in our intense work culture, especially in the tech industry. However, it does not justify the unhealthy nature of overworking. You might say, working hard is a merit, why would you pathologize the positive work ethics rather than promote it? Yes, when you have too much of a good thing, it starts to create more serious troubles. Let’s take a look at the rest of the symptoms.

II. Distress

• Use despite physical or psychological problems caused by use

Over the years, my physical and mental wellness troubles started to occur more frequently. Stress caused pain in the body, stomachaches and headaches, as well as sleeping problems. Occasional burnouts crushed my mental health. However, they did not prompt me to change my overworking patterns. In fact, I began to see work as a remedy, which seems to be effective in numbing myself, at least temporarily. As long as I indulge in the work and get the achievement I wanted, I don’t have to face those problems directly.

• Craving

People with addictive disorders have this chronic craving to use the substance. I would consider this feeling too familiar as I crave working when I’m not. I’m available on chat/slack 24/7; I seep into the home office when kids are not around during the weekend. Even when walking my dog, I can’t stand still in the park without checking my work email on the phone. Initially craving is targeted towards the final outcome for an achievement, you really look for the feeling of attaining that success. Naturally that craving gets transformed to work itself that you forgot the goalpost from time to time. You simply don’t feel yourself if you are not working.

• Tolerance

Earlier in a person’s career, success seeking is a positive reinforcement. I get better at my work, and I obtain more achievements. I proved myself that I could be a great engineer in Silicon Valley, then I started my own company and achieved profitability with millions of dollars of revenue. It all feels good until you get used to it. Gradually and unconsciously, you up your game. I had to start a second company to be greater, raising more funding from more prominent VCs. Before long you start to play in higher and higher leagues. I went on to prove that I could be a great product manager in prestige tech companies, building teams and launching coolest AI products. Certain achievements you would be dreaming about just a few years ago don’t feel exciting any more. You develop tolerance. The only way of gaining more satisfaction is to work harder and to achieve more success. Happiness is fleeting and unguaranteed; stress and dissatisfaction is the constant.

• Withdrawal

Withdrawal syndromes in substance abuse disorders can be found when a person stopped or reduced the use for a prolonged period. They would feel anxious, agitated, or unease, all the way to more serious psychological or physiological symptoms. I found myself in similar mode when I’m not working on accomplishment-related activities or finding myself lacking forward momentum for some prolonged period of time. My mood would get very low, sometimes with emotional disturbance. And then the craving kicks in, which starts another cycle.

Reflecting on my journey in developing this disorder, I had a lot of the above symptoms earlier in my career. They also seem to be on a very personal level. Even though I had the subconscious idea that something was not right, the problems were still constrained to myself. I was very young, and no matter what I could push forward with my energy and willpower. Things started to really go south when other parts of my life broke apart.

III. Dysfunction (in other areas of life)

• Fail to fulfill major roles in other areas of life

During the years of building my two startups, I became a father of two young kids. In my twenties, I heard people caution that raising kids is a lot of hard work. Oh yes of course, I love hard work. Boy, I was so naive. The burden of running startups and being a parent really brought me to my knees. Unfortunately, being the kind of addict I was, I could not allow myself to drop the balls at work. Majority of the kids’ duty fell on my wife who also had a booming career. The family dynamics were a mess and I was in constant self-blame in not being the husband and father I always wanted to be.

• Persistent social or interpersonal problems caused by use, with important activities given up or reduced

When my workaholism is in overdrive, everything else gets moved to the back burner. In retrospection, during the months that work stress took over me emotionally, I was indeed like a walking zombie at home. None of my friends would see me for a prolonged period of time because I was either working or sleeping. There was no weekday or weekend, not even day or night.

Over-enthusiasm, over-work and intense suffering. After cycling through this familiar and vicious pattern for more than a decade, at the wake of another burnout, I suddenly came to realize that this was not only unhealthy and counterproductive but also the root cause of so many of my life problems. I could not not change. Over the last two years, I embarked on a long journey to come off this addiction. In the next post, I’ll share some key learnings from my recovery and how past sufferings could become the foundation for a more fulfilling life going forward.

I couldn’t have been more wrong.

A few years ago, I found that my self-talk seemed to ebb and flow correlating to my fluctuating stress levels. I started to examine deeply what really happens in my mind; and was surprised to find out that the majority of the self-talk is not only wasteful but also harmful. Unconsciously, the content was full of negative judgements on myself, others and any situation I face. It also contains repeating reinforcement of my existing beliefs, self-doubt and blame, as well as tons of ruminating worry about the unknown future. If it is contributing to anything, it is more to my stress and anxiety than to my wellbeing. I also came to realize that almost everyone does this unproductive self-talk more often than they would like. Especially when we’re facing challenging situations or experiencing difficult emotions, these voices get louder and very few of them would turn out to be constructive, not to mention help us grow.

My goal is to be able to get out of my self-talk, especially when it starts to ruminate and gets stuck in a negative loop, and to refocus my energy on constructive actions with a more clear mind. Unconsciously, when the voice in my head gets louder and noisier, guess what I do? Yes, I do more self-talk. It’s not unlike a computer application that runs on overdrive, taking up more and more chunks of the CPU, memory and network resources, but not producing much useful output.

However, as a mental habit that I formed way earlier in my life, I found it extremely hard to unlearn. We as humans have very limited capacity to radically transform thinking through our own self-talk. With this awareness, I started to view my self-talk as an addiction and I took my recovery seriously. I tried almost all methods, and here are a few that I want to highlight in this post. I’d like to clarify that these methods are not generalized wellness tips or suggestions; instead they are targeted to solve the specific problems of self-talk. They proved to work very well for me; but every brain works differently, so your mileage might vary.

Writing (the right kind of words)

When you put your thoughts down on paper or computer in a self-reflective way, the act has a special effect on brain functioning. The key here is to be honest, transparent, and diligent in reflecting very deep into your thoughts and emotions in your writing. When your self-talk starts to ruminate, it’s like running in a traffic circle. Writing provides an offramp that guides the thoughts out of the loop. The first effect is it offloads the burden on the mind, like freeing up the computing resource that the self-talk program has been taking up.

Secondly, it also serves to enable you to observe and process the content in your mind. When you are wearing a hat on your head, you can’t see what the hat looks like, until you take it down. Similarly, when your thoughts are running in overdrive in your head, you can’t clearly observe what they are, until you get them out.

• Pros: Flexible with little restriction. No cost, and self-paced treatment.

• Cons: Not everyone has a habit of writing or enjoys doing it. It takes practice to master the type of writing that works for you. It also costs quite some time to do it, and needs self-discipline to take effect.

Reading (the right kind of books)

When I was in the middle of a self-talk episode, I found reading to be of good relief. The type of book is the key in its effect on healing. I used to like fiction but they are not helpful to me in treating self-talk. Reading fiction, in its effect on my brain, is like watching movies or TV shows. It gets my brain fully distracted on something else, but only temporarily. After I put down the book or turn off the TV, the self-talk comes back with full energy, as if it just took a nice break. A good type of books to combat self-talk would be non-fiction that is tangentially related to the trouble that is bothering you. For example, when I was contemplating my career transition a few years ago, I found Arthur Brooks’ book From Strength to Strength very inspiring and it helped me sort through some conflicting thoughts. The Conscious Parent by Shefali Tsabary gave me the awareness to get out of my parenting struggles. It really depends on the problem you’re working on so there is no rule in exactly what book would help you.

Pro tip: Ask ChatGPT/Gemini for book recommendations by sharing some of your reflective writing could turn out to be useful.

• Pros: Flexible with little restriction. Low cost, and self-paced treatment.

• Cons: Finding the right kind of books that fit your needs could be hard. Benefits may not be immediate or directly observable.

Talking (with the right kind of person)

Ok, let’s be honest, transforming our thinking is hard. Even though we do have quite some methods to help ourselves, often we don’t change well on our own. As Daniel Kahneman said in his book Thinking, Fast and Slow, we resist self-exploration especially when emotions are involved. To stop the adverse patterns in our self-talk, someone outside our head needs to disrupt our thinking by reflecting our thoughts back to us and asking questions that prompt us to wonder why we think the way we do. Talking with someone such as a coach serves this purpose. An effective conversation should involve reflective statements and thought-provoking questions that enable us to see our concocted stories as if they were laid out in front of us in a book to be read and analyzed.

The key is to choose the right kind of person to have this conversation. As I mentioned in a previous post, a friend could help in some cases but usually they are not effective in dealing with your self-talk. In my personal experience, having these types of conversations made the most long-lasting changes in my relationship with my own mind.

• Pros: Powerful and effective, with long-term benefits of growth. Personalized solution.

• Cons: Finding the right person who fits your needs and can adopt the coaching method. Associated cost and less flexibility.

In essence, when our self-talk stops serving us constructively and becomes a mental burden, we need to redirect and expand our mind. Adults need this help as much as children do, and sometimes more. As we age, we become more rigid in our thinking. We become masters at rationalizing our actions, ignoring our emotions, and finding what confirms our beliefs. As a self-talk addict in recovery, I tried and collected a full array of tools to be useful and effective in treating different types of issues. Apart from the above, I’ll share more in future posts about the effects of meditation, physical exercises on mental wellness. Stay tuned.

As I alluded to in a previous post, coaching and therapy share quite a lot of philosophical underpinnings, psychological theories and intervention techniques. In my master’s training of counseling at New York University’s Applied Psychology department, we learned a lot of psychological foundations of human behavior, theories of various psychotherapy approaches, and almost covered all the popular intervention techniques. When I had my coaching training, one of the core competencies from International Coaching Federation (ICF) is for a coach to know when and how to refer to a therapist when appropriate.

To elaborate on their distinctions and correlation, I was hesitant to use an illustration because it will surely be an oversimplification. Nothing visual could justify the complexity of working with the human mind; but I still tried to do the impossible.

Granted, a picture is worth a thousand words; but please bear with me as I have to add a few more words.

What’s the difference?

As illustrated in the diagram above, there are two dimensions, the functional spectrum (vertical axis) and the temporal spectrum (horizontal axis). It’s gradient on both these two spectrums; and the green shades are therapy’s focus while the blue shades are coaching’s focus.

Functional spectrum. If a client is more functioning in terms of their daily work and life, free of psychological disorders, they would benefit more from coaching. If an individual is distressed with mental illness with dysfunction of their life, they would benefit more from therapy. It’s easier to conceptualize this distinction by the extremes. On the top end of the spectrum, you might imagine people like a newly-promoted CEO of a Fortune 500 company, who works with their coach on preparing to take on the new job and turn the company’s performance around. On the bottom end of the spectrum, you could find veterans suffering from post-traumatic stress disorder who are so debilitated by the symptoms that they are not able to leave their home. They would benefit from therapy work to better process the trauma and cope with these devastating emotional reactions.

Temporal spectrum. Coaching typically focuses on the present and future. A coaching session would oftentimes discuss visioning and success, addressing the obstacles in the present moment while moving into the future. Therapy, relatively speaking, emphasizes more on the past in order to understand the present. Bringing the two examples above, coaching the CEO focuses on the current issues around performance and development and aims to achieve the goals towards the future; while the veteran would benefit from delving into his past trauma and the deep-seated emotional issues around it in order to enable healing and recovery.

With all of the above said, what’s really useful to think about is the vast spectrum in the middle.

Either Or? Both?

Human beings are always more complex than any categorical mental model. Highly functioning individuals might also suffer from mental health and could use therapy help; so on the functional spectrum, the line could be blurred. A coaching client would also benefit from understanding their past to uncover the root of their limiting thoughts in order to move forward with their goals; so on the temporal spectrum it’s also not entirely black and white.

Apart from the extreme ends of these spectrums, most individuals live in the middle of the diagram. In a lot of cases, people could actually benefit from both coaching and therapy. In some cases, therapy and coaching can happen in sequence, for example, a few months of therapy and then transition to coaching. Oftentimes coaching and therapy can occur during the same period of time while focusing on different aspects and complement each other in accelerating the progress. Here are a few client examples.

Sam has a PhD in biochemistry, is the author of dozens of patents, and head of R&D in a biotech company. Despite his various accomplishments, he always feels not enough, and believes that he could have done better and should work much harder to achieve more success.

Coaching

Sam: I remember when I was 10 years old and got a B on my report card, my father was so furious and disciplined me for that miss.
Coach: Oh I’m sorry to hear that. It sounds like your father definitely had very high expectations of your academic accomplishment as a child. Now as an adult, what choice would you make about your accomplishment?
[The following conversation could focus on how a previous belief structure manifest in the present, bring awareness and potential change.]

Therapy

Sam: I remember when I was 10 years old and got a B on my report card, my father was so furious and disciplined me for that miss.

Therapy: Oh I’m sorry to hear that. It sounds like a traumatic experience for you. Tell me more about that.
[The following conversation could focus on processing that memory and experience, construct a better relationship and narrative around it.]

---

• Lauren has been struggling with binge eating problems for years and was not able to control her weight. She suffered an abusive childhood, and has been using eating and obesity to feel safer.

Coaching

Lauren: Everytime I think about all the diet and exercise efforts I need to take in order to bring my weight down, it’s just too much and horrifies me.

Coach: The efforts surely sound daunting when they are perceived as a big chunk. We can try breaking it down to make them more manageable. What would be some immediate things you could do in the next week to get started?
[The following conversation could focus on chunking down the acion tasks and motivate the client to take small steps towards a big goal.]

Therapy

Lauren: I hate how I look in the mirror; but I’m also fearful if I lose weight.

Therapy: I understand that you started to have the eating problem in middle school. What was your memory and experience about your weight and appearance back then?
[The following conversation could focus on understanding the past and how it builds an unconscious and limiting belief about her weight.]

---

In therapy, one of the major goals is to heal a person. Catching on an emotional event and diving deep into processing the past is a key element. In coaching, on the other hand, we choose to briefly dip into emotions, with which we use as a cue to focus our position in the present and our path for the future.

In essence, knowing what type of resources would be beneficial to an individual shouldn’t reside solely within the coach’s toolkit. It’s a vital piece of knowledge that empowers the client, too. The path to self-discovery and a fulfilling life is a dynamic journey, with each individual at different phases and requiring unique forms of support. As a coach, my deepest satisfaction lies in acting as a dedicated partner and companion, providing and suggesting the tailored resources necessary to illuminate and ease the steps of their remarkable journey.

“Who doesn’t?” I told myself, “You have to use feedback to learn and grow. If you feel bad about it, then it must be your problem. Your goal is to outgrow it.”

I grew and grew over the years. As a startup founder, I solicited feedback from various stakeholders, my investors, advisors, employees, and our customers. I learned how to take negative feedback, even if unexpected or unjustified, and try to turn it into a constructive outcome. In my corporate job at Amazon and Google, feedback is built into the annual/quarterly evaluation cycle (called Forte at Amazon and Perf/GRAD at Google). We also encourage managers and peers to give more timely, constructive feedback to help people improve and grow.

Feedback, especially negative ones, seems to serve a really positive intention. Truth is, after so many years, I still don’t like it. Having been an employee, founder and leader myself, I have also realized that most people don’t. Research shows that even though managers believe their employees are open to constructive feedback; the employees report that despite their willingness to grow, they do not like it.

Then what’s the problem?

We do want to learn and grow, but our human brain is also hardwired with a more basic need: to be accepted. In the article “Find the Coaching in Criticism,” Harvard Law professors Sheila Heen and Douglas Stone found that even well-intentioned feedback “spark an emotional reaction, inject tension into the relationship, and bring communication to a halt”. Negative feedback, especially if unsolicited, is painful.

We do want to learn and grow, but our human brain is also hardwired with a more basic need: to be accepted. In the article “Find the Coaching in Criticism,” Harvard Law professors Sheila Heen and Douglas Stone found that even well-intentioned feedback “spark an emotional reaction, inject tension into the relationship, and bring communication to a halt”. Negative feedback, especially if unsolicited, is painful.

Consider the following scenario in the workplace between Evan and his manager Peter:

Evan: Jeff is really hard to work with. I was hoping to move our project forward by chatting with him through the messages but he insisted on having a live meeting. This project is very important and we can’t delay it any more. I was already overwhelmed with my schedules and really don’t have time to do that.

Peter: I appreciate you taking on a lot of tasks at the same time; but I would like to give you a piece of feedback as I have noticed it as a recurring pattern recently: You need to better manage your priority. If this project is really important, you should just meet with Jeff.

For Evan, the hard part of this conversation is that what Peter said is not entirely wrong, in reality the feedback and suggestion could be totally valid given the situation. This fact makes the feedback emotionally difficult for him to accept. After thinking it through and processing the reasoning behind it, Evan might feel very defeated and realized that he made a huge mistake and was losing trust from his manager. Granted, there are situations that a manager might intentionally want to provide feedback in this way to push their employees to make necessary changes. However, these conversations could turn out to be less effective than people had expected, and cause long-term harm.

What’s a better way?

In a lot of scenarios, there is a more productive approach to help people uncover blindspots, facilitate learning and growth. Leveraging the coaching approach of reflective inquiries, one person can help the other engage in reflecting their own thinking, examining the beliefs and identifying gaps. Through this partnership, insights will reveal, action steps emerge, and growth accelerates. At the end of the day, we are more critical to ourselves than anyone else, don’t we? We are also experts in knowing our problems, making decisions and changing ourselves.

Consider how the same conversation above could go in a different direction:

Evan: Jeff is really hard to work with. I was hoping to move our project forward by chatting through the messages but he insisted on having a live meeting. I was already overwhelmed with my schedules and really don’t have time to do that.

Peter: I appreciate you taking on a lot of important tasks at the same time; and you seemed really frustrated about his meeting request. What was making it particularly hard for you?

Evan: Hmm… I guess I was spreading myself too thin recently. The amount of work and deadlines have been making me swamped. I surely need to find a better way to prioritize.

Peter: Sounds like this project is definitely one of the highest priorities but is now blocked. Given this situation, what do you think is the best way to move it forward?

Evan: I guess the best option right now is to have a quick meeting with him to unblock the next steps, and at the same time set up the expectation that we could use messages in the future to drive efficiency.

Peter: Sounds like a great plan. Handling too many pressing projects could definitely be challenging and may be counterproductive. Let’s talk about how we could shift the tasks a bit and help you better prioritize.

In this example, Peter used the coaching approach of reflective inquiry to engage in a collaboration conversation with Evan. Through the thought process, Peter elicited the solution for the pressing problem, as well as a broader growth area of better prioritization. The best part of it is that all the insights were from Evan himself, which makes it much easier for him to adopt in their future behaviors.

To be fair, this technique is not always suitable for all circumstances. However, as executives, managers and leaders in organizations, leveraging coaching as an alternative to direct feedback certainly contributes to more powerful communication. Used effectively in management, it could drive more effective change, boost morale, and foster long-term professional growth for our teams.

The New Age of Entrepreneurship Is Here

For the past year, the conversation around AI has been dominated by one theme: job loss.

It's a narrative of automation and displacement, and much of it is centered on fear. While those concerns are valid, they miss the bigger picture. The same technology that's changing the nature of work is also launching a new golden age of entrepreneurship.

This isn't just a minor shift. It's a fundamental change in how businesses are created, who gets to build them, and how fast they can grow. The barrier to entry for starting a tech company is collapsing, and a new class of founders is emerging.

Instead of fearing a future with fewer jobs, we should be building a future with more employers.

1. Your New Co-Founder is an AI 🤖

In the past, one of the biggest hurdles to starting a tech company was the technology itself. If you weren't a software engineer, you had to find one, and then build a large team of them. This takes time, money, equity, and tons of risk associated. That era is ending. AI has made the basic skills of creation accessible to more people, acting as a powerful partner for anyone with a strong idea.

• Productivity and Research: A single founder can now handle work that once required a founding team. AI tools can automate routine tasks like data entry and scheduling, analyze market trends, and draft research summaries in minutes instead of days or weeks.

• Design and Development: You no longer need to be a coding expert to build a working product. AI-powered platforms can generate designs, create website prototypes, and write and fix code. In fact, a quarter of startups in a recent Y Combinator group had codebases that were almost entirely AI-generated. Tools like Cursor and the JetBrains AI Assistant act as partners in the development process, dramatically speeding up the time it takes to get a product to market.

AI doesn't replace the founder's vision. It clears away the tedious and technical tasks, allowing entrepreneurs to focus on what truly matters: solving customer problems.

This means the barrier to starting has never been lower. The barrier to winning is now about focus and the speed of your learning loop.

2. The New Startup Model: The Rise of the Lean Unicorn 🦄

The old startup playbook is also becoming outdated. The traditional model was: raise money, hire a large team, burn cash to grow, and repeat. This approach was expensive and created complex organizations.

AI is creating a new economic model. The new way is: raise what you need, invest in powerful tools and a small, highly-skilled team, and automate the rest.

We're now seeing the rise of the "40-person unicorn," and that number keeps shrinking. Companies are reaching huge revenue milestones with surprisingly small teams.

• Highly Leveraged Teams: An AI-native company uses smart systems to handle sales, marketing, and customer support. AI agents can find new customers, send personalized emails, and answer over 80% of support questions without a human stepping in. This means a small team can achieve the output of a much larger one.

• Smarter Use of Capital: This lean approach leads to incredible capital efficiency. AI-native startups are showing revenue per employee that is 4 to 5 times higher than typical software companies. In today's economic climate, investors value this efficiency more than ever. It allows startups to grow sustainably, secure better valuations, and lets founders keep more of their company.

3. The Democratization of the Entrepreneurship ✊

Perhaps the most significant shift is in who gets to be a founder. For decades, tech entrepreneurship was largely the domain of those with deep software development or computer science backgrounds. That's no longer the case.

AI is leveling the playing field in two important ways:

• Domain Expertise is the New Edge: AI isn't just for software companies; it's rewiring operations in traditional industries like law, healthcare, and manufacturing. A lawyer, doctor, or supply chain expert with deep industry knowledge now has a massive advantage. They know the real-world pain points, workflows, and data that are essential for building a valuable AI solution. They can use AI-assisted coding and low-code tools to build products without needing to be a veteran programmer. The contrarian view is to not start with a grand "AI platform," but with a single, high-value workflow and become the best in the world at solving that specific problem.

• The Learning Curve is Reset: In the age of AI, everyone is a beginner again. A recent college graduate and a 20-year industry veteran are starting from a similar point on the learning curve for this new model of building. The advantage goes to those who can learn and adapt the fastest, not those with the most established credentials. This opens the door for a new generation of founders from all backgrounds and walks of life.

This wave is also enabling the rise of "nanobusinesses"—small, highly focused companies that serve niche markets. Because AI lowers the cost of reaching customers and building products, a single person can now identify a specific problem, build a solution, and market it to a global audience from their laptop.

4. Capital Efficiency is Your New Superpower 💰

In this new era, valuations are rewarding efficient growth. Startups that can achieve significant revenue with a small team are being valued at a premium. We are seeing companies with fewer than 100 employees reaching nine-figure annual recurring revenue (ARR).

Founders should now obsess over metrics like ARR per full-time employee (FTE), aiming to surpass $250k early on. Keeping the team lean and the burn rate tight is no longer just about survival; it’s about optionality. It gives you more control over your company, allows you to negotiate better terms with investors, and ultimately lets you keep more of the business you're building.

A Call to Action: Build the Future 🚀

The story of AI will be one of transformation. Some jobs will change, and new skills will be required. Ultimately, automation frees up human capacity to focus on what we do best: judgment, creativity, and building relationships.

The number of entrepreneurs in the world has always been a small fraction of the population, but in the AI era, that number is set to grow. This is the moment to embrace the next generation of entrepreneurship. It's a future with more founders, more small and nimble companies, and more specialized jobs that leverage uniquely human skills.

The question is no longer if you can build your idea. The only question is if you will.

Introduction

For decades, technologists and product leaders are forced to make a painful choice between three things:

• Scale,

• Customization, and

• Quality.

When you are building technology solutions, whether it’s software, hardware or services, you could only pick two. Never all three.

This was the iron law of technology strategy, a constraint that shaped the very structure of our products, our companies, and our markets.

In the age of AI, things are different. I call it The AI Trilemma Advantage.

For the first time ever, technology isn't just bending the old rules of business. It's breaking them. AI resolves this trilemma, creating a massive new advantage for those who get it.

This article gives you a new way to think about that strategy.

This isn't just a theory. It’s a guide to the new physics of building in the AI era, and a playbook that could help you in defining your AI product positioning and building towards defensible moat.

Let's get started.

Section 1: The Old Rules: The Prison of "Pick Two"

To understand why AI is such a big deal, you first need to understand the box we’ve all been trapped in. The old tech trilemma is that you could not have scale, customization, and quality all at once. It was structurally impossible. Every founder and product builder had to make a compromise.

This led to three basic business models:

• Scale + Quality (No/Low Customization): Think of the Ford Model T. You got a great, reliable car that was affordable for millions. The catch? You could have any color as long as it was black. This was the world of one-size-fits-all software. Effective, but generic.

• Quality + Customization (No/Low Scale): Think of a personal tutor or a bespoke suit. The quality is perfect, and it’s tailored just for you. The problem? It doesn’t scale. It’s expensive and only available to a few. High-end consulting and custom software projects live here.

• Scale + Customization (No/Low Quality): This was the first clumsy attempt to use tech for personalization. Remember those marketing emails that just inserted your first name? That’s it. It was personalization at scale, but it was shallow, rigid, and low-quality. The tech was just following simple rules. Quite a few AI products and companies from previous waves of AI hype cycles are like this. They put in tons of “bandaid” rules in the system, but the quality never meets the bar.

This wasn't a choice; it was a technological prison.

Why? First, the Human Expertise Bottleneck. Real quality and customization required a smart human, and you can’t clone humans. Second, Computational Rigidity. Old software was brittle. It followed rules. It couldn't handle nuance or adapt on the fly. Third, Information Overload. The complexity of trying to understand millions of individual users and maintain quality was just too much for pre-AI systems to handle.

These limits defined the strategy for every company. This table lays out the old world.

Table 1: The Traditional Technology Trade-Off Matrix

Section 2: The Resolution: How AI Tackles the Impossible

AI isn’t just a faster horse. It’s a different kind of engine altogether. It doesn’t just bend the old rules; it breaks them by attacking the core constraints of the trilemma. This happens for three reasons.

First, Scalable Intelligence. For the first time, we can scale cognitive work, which used to be found only inside the human brain. AI models can reason, create, and make judgments at near-zero marginal cost. This shatters the human expertise bottleneck.

Second, Dynamic Contextual Understanding. Old software was rigid. On the contrary, AI understands natural language, nuance, and context. This allows for meaningful, adaptive personalization at scale, not the robotic rules and patterns we had before.

Third, Continuous Learning. AI systems get smarter over time. More data and more users don't degrade the system; they improve it. This creates a virtuous cycle where scale actually improves customization and quality.

When AI resolves the trilemma, it redefines what each pillar means :

• Scalability becomes Infinite. The goal isn't just serving millions; it's serving billions, with each user making the system smarter.

• Customization becomes Hyper-Personalization. We move from basic segments to true 1:1 experiences that feel like the product was built just for you.

• Quality becomes Superhuman Performance. The goal is no longer to match a human expert but to exceed them, with perfect consistency, every single time.

This convergence is what I call the Trilemma Advantage: a durable competitive edge for companies that build their entire business around delivering all three.

This isn't just about technology or product; it's a new economic model.

Old companies needed huge sales and support teams to deliver custom experiences at scale. AI automates cognitive work. This is why we're seeing the rise of the "40-person unicorn"—companies achieving massive scale and valuation without the massive headcount. The Trilemma Advantage is a financial weapon.

Section 3: The Trilemma Advantage in Action

This isn't just a theory. The frontier companies in various industries are already using the Trilemma Advantage to build incredible AI products and competitive edges.

3.1 Education: A Personal Tutor for Every Child

Education has always been stuck in the trilemma. You could have a great 1:1 human tutor (Quality + Customization) or a mass-market textbook (Scale + Quality). AI is finally breaking this trade-off by delivering the equivalent of a personal tutor to every student on earth.

• Khan Academy's Khanmigo: Khanmigo is an AI-powered tutor and teaching assistant built on GPT-4 and trained on Khan Academy's world-class content library. It perfectly demonstrates the resolution of the trilemma. It provides

• Quality by using a Socratic method that guides students to answers through critical thinking rather than just providing them. It delivers deep

• Customization by adapting to a student's interests—framing math problems around sports, for example—and acting as a "thinking partner" that offers layered hints based on their specific struggles. And it achieves massive

• Scale, with pilot programs running in hundreds of school districts, making this high-quality, personalized experience available to thousands of students simultaneously. The results are tangible: in one pilot school, after just one semester of using Khanmigo for geometry, there were no students failing the class.

3.2 Healthcare: A Specialist in Every Clinic

In healthcare, the trilemma can be a matter of life and death. Access to a top pathologist or a rapid stroke response team has always been limited by human availability (Quality + Customization without Scale). AI is democratizing this expertise.

• PathAI & Viz.ai: PathAI uses machine learning to help pathologists make faster, more accurate cancer diagnoses, delivering specialist-level Quality and Customization to labs everywhere (Scale). Viz.ai uses AI to detect strokes from scans and instantly coordinates the hospital care team on their phones. It’s used in over 1,700 hospitals and has been shown to cut the time it takes to transfer critical patients nearly in half.

3.3 Commerce: A Bespoke Store for a Billion People

Digital commerce is where the Trilemma Advantage is most obvious. The old choice was between a personal shopper (custom but not scalable) and a department store (scalable but generic).

• Netflix, Amazon & Starbucks: These companies use AI to deliver hyper-customized experiences at a massive scale. The quality is measured in dollars. Netflix’s recommendation engine saves it $1 billion a year in reduced churn. An estimated 35% of Amazon’s revenue comes from its recommendation engine. Starbucks’ AI platform led to a 30% increase in marketing ROI. They turned a huge catalog and user base into a strategic weapon.

3.4 Legal Services Automated: The Partner in Every Laptop

The legal world has long been the poster child for the "Quality + Customization without Scale" model. A top lawyer provides exceptional, tailored advice, but their time is finite and incredibly expensive. AI is now productizing that expertise.

AI Contract Review: Companies like Harvey, Ironclad, and Sirion are deploying AI that acts like a superhuman paralegal. These platforms can review hundreds of contracts in seconds (Scale), check them against a company's specific legal playbook for compliance (Customization), and flag risky clauses with a precision that reduces human error (Quality). The impact is staggering. One analysis found AI could reduce the cost of a legal contract review by 99.97%. Another report noted that AI allows legal teams to review six times the number of contracts, turning a cost center into a strategic advantage.

Conclusion: Stop Building Features, Start Building Architectures

So, what’s the takeaway? The old rules are dead. The trade-offs that defined your strategy for the last 20 years are gone. AI resolves the trilemma of Scale, Customization, and Quality. We've seen how it works and how the best companies are using it. We've laid out a playbook for you to do the same.

The job of a product leader today is to stop being a builder of AI features and start being an architect of a Trilemma-Native business. This means building your product, and even your economic model around scalable intelligence. It means being lean and fast where incumbents are fat and slow. And it means seeing data not as a pile of stuff to own, but as the fuel for a learning engine that never stops. This is how you win.

Product leadership is not for the faint of heart. The job is to embrace ambiguity and chart a course to a future that doesn't exist yet. You will make mistakes. The question isn't if, but how we can learn from them and avoid similar ones.

I’ve been in the arena for a while now—leading teams at Google and Amazon, founding two AI-driven startups, and seeing products succeed and fail spectacularly. My mistakes have cost lots of money, engineering effort and user trust, and, most painfully, led to letting go of people I cared about.

This isn't a confession. It's for our learning. I'm sharing my biggest mistakes to give you a cheat sheet, with some of the most deceptive traps that look like great opportunities. Recognize them, so you don't have to learn the hard way like I did.

Mistake #1: Chasing Shiny Objects (The Siren Song of "Synergy")

In my first startup, we built purchasing data analytics and marketing solutions for large consumer brands and financial institutions. We were two years in, hitting product-market fit, and seeing healthy revenue growth. We were focused and disciplined.

Then the shiny object appeared: mobile. Smartphones were taking over the world. Consumer apps were seeing exponential growth. I became convinced we needed a mobile strategy. I sold myself and my board on a vision: a consumer-facing mobile app for credit card users that would have "synergy" with our core enterprise business.

The Mistake: "Synergy" is one of the most dangerous words in business. It's often an excuse to lose focus.

I made a bold, and ultimately catastrophic, decision. I hired a new team of over 20 designers and engineers. We built a completely different product on a different tech stack for a different customer with a different go-to-market model. We were now two different companies under one roof.

Six months later, reality hit me like a ton of bricks. We didn't have the resources or the bandwidth to fight a war on two fronts. Our core business was suffering from the distraction, and the mobile app was a money pit.

I had to make the most painful decision of my career: I shut down the consumer division and laid off the entire team. I will never forget the looks on their faces. That mistake wasn't just financial; it was a deep emotional blow to our culture and morale.

How to Avoid This Trap:

• Be the Best in the World, or Don't Bother. Before starting a new initiative, apply this brutal filter: "Are we determined to become, without question, one of the best in the world at this?" If the answer is anything less than a resounding "yes," don't do it. Being mediocre at two things is a recipe for failure.

• Define Your Anti-Roadmap. A great strategy is defined by what you choose not to do. Create an "anti-roadmap"—a shared, explicit list of the markets, features, and customers you will ignore. This creates ruthless focus and makes it easy to say no to shiny objects.

• Focus is a Verb. Focus isn't a poster on the wall. It's the daily, painful act of saying no to good ideas so you can execute on the great ones. It's about allocating 100% of your best resources to your #1 priority, not 10% to ten different priorities.

Mistake #2: Building on a Foundation of Unproven Tech

Googlers are dreamers. That’s what attracted me to join the company as a product leader. I want to build amazing, almost impossible, products leveraging cutting-edge technologies.

My first role at Google is at Nest, where I led the product team to define and build the future of intelligent smart home experience. We were bold and provocative: a truly proactive smart home system that anticipates your every need and takes actions on behalf of you and your family. The problem is, that vision requires a dozen underlying technologies to work perfectly in concert—sensor fusion, on-device processing, predictive models, audio and visual recognition, and more. The even more serious bottleneck is that users expect a very high level of certainty and reliability in their home, before they could be ready to adopt that proactivity.

After months and months of efforts in designing, prototyping and testing, we ended up winding down the initiative.

The Mistake: Committing to a product roadmap before the key enabling technologies are mature and reliable, especially for a consumer experience that reliability and control is critical (home experience).

How to Avoid This Trap:

• Fall in love with feasibility, not just possibility. Innovation without pragmatism is self-sabotage.

• Isolate and Attack the Riskiest Assumptions. Don't build the whole system at once. Identify the single biggest technical leap of faith you are making. Run a focused, time-boxed R&D sprint to prove or disprove that one assumption. This iterative, science-based approach to de-risking is how you separate science fiction from viable products.

Mistake #3: Over-indexing on a Single "Perfect" Customer

My second startup builds AI chatbot solutions for enterprise customer service. This was pre-LLM, when building a truly smart conversational agent was a monumental task. We were a small, hungry team of five engineers when we landed the dream design partner: Lyft.

They were a rocket ship in their boom days, and we were thrilled to be along for the ride. I dove in headfirst, embedding myself with their team. We lived and breathed their problems. We built our entire V1 product to solve Lyft's specific, complex use case for driver and rider support. And it worked. The launch was a massive success. Lyft became our biggest paying customer and a powerful logo that opened doors.

The Mistake: We hadn't built a real, scalable product. We had built a custom solution for one client. We were, in effect, a highly efficient, outsourced engineering team for our largest customer.

When we went to sell to our next ten customers in e-commerce, healthcare, and finance, we hit a wall. Their problems were different. Their scale was different. Their data was different. We discovered, to our surprise, that Lyft's use case was an outlier. We had optimized for the exception, not the rule. The "traction" we thought we had was a mirage.

The next 18 months were a painful process of "unlearning", and redeveloping. We had to rip out hard-coded logic, re-architect our data models, and fundamentally reposition the product. We had to tell new customers, "No, we can't do that yet," while our engineers worked furiously to generalize a platform we had built to be specific. Fortunately we gradually expanded our features and gained traction from other customers.

How to Avoid This Trap:

• The Rule of Three. Never commit to a core product direction until you have validated the underlying problem with at least three unaffiliated, potential customers who fit your ideal profile. One customer is an anecdote. Two is a coincidence. Three is a pattern.

• Abstract the Problem. When a design partner requests a feature, your job is not to just build it. It's to ask, "What is the universal problem this specific request is a symptom of?" If they ask for a "red button that does X," you need to understand the job-to-be-done behind the button. Build for the job, not the button. I know first-hand that it’s easier said than done, especially when you don’t have much leverage in the beginning.

• Contractually Define the Relationship. A design partner is not a typical customer. The contract should reflect this. They get early access and influence, but in exchange, they commit to providing feedback on a generalizable solution. This frames the conversation correctly from day one. You're building a product with them, not for them.

Mistake #4: Believing You Can Force a User Habit

When I was at Amazon, I led discovery and engagement for Alexa. The Echo was the hottest gadget on the planet in those years. Everyone and their grandma has one on their kitchen counter or nightstand. We had a mandate: make Alexa an indispensable daily habit.

So we rolled our sleeves and shipped. We built features to "educate" users, prompted them with suggestions, and put "skills" on the device home screen. We relentlessly pushed all the amazing things Alexa could do, convinced that if users just knew about them and tried them enough times, they'd be hooked. Our adoption and engagement numbers were indeed amazing. But we were so focused on them that we forgot the cardinal rule: you can't manufacture a habit.

The Mistake: We were solving our organization’s problem, not the user's. Our metrics were focused too much on engagement, not enough on value. The result? We annoyed people. Quite a lot.

The infamous "By the way..." feature became a meme for a reason. Users ask for the weather and then get a lecture about listening to music. It broke user trust more often than we wished. A habit is the byproduct of delivering indispensable value, so much so that the user pulls the product into their life. We were pushing. We had to reverse a lot of efforts later on and rebalance the value we deliver and the growth efforts we make.

How to Avoid This Trap:

• Solve for "Pull," Not "Push." Stop asking, "How can we get users to engage more?" and start asking, "What problem can we solve so masterfully that users can't imagine going back?" A great product doesn't need to nag. Its value is its marketing.

• Measure Task Success, Not Just Engagement. Are users successfully completing their goals? Are they solving their problems faster or better than before? A user who opens your app once a week to solve a critical problem is infinitely more valuable than one who opens it daily out of annoyance and closes it immediately.

• Practice "Quiet Confidence." A truly great product is confident in its core value. It doesn't need to constantly advertise its secondary features. Let users discover them organically through contextual, non-intrusive cues. When a user is already doing X, that might be the right moment to gently suggest Y. Don't shout at them when they walk in the door.

Mistake #5: Managing AI Products with the Old Playbook

When I switched teams within Google and led product efforts in launching Gemini, I saw firsthand how even the smartest, most experienced product managers were struggling to design and build AI products. For decades, we as product leaders were trained to build deterministic systems. You write a spec, the engineers build it, and we verify it by testing that the feature behaves predictably.

The Mistake: Applying this feature-driven, deterministic playbook to probabilistic systems like modern AI.

Modern AI is not deterministic. It's a core feature, not a bug. Demanding certainty, fixed timelines, and predictable roadmaps from a system designed to operate on probability is a recipe for disaster. It leads to brittle products that feel like clunky flowcharts, not intelligent agents. It leads to frustrated teams and missed deadlines. As I've written before, the failure rate for AI projects is over 80%, and this mindset is the primary culprit.

How to Avoid This Trap:

• Make the Eval Framework the Product Spec. In the AI era, your most important document is not the PRD. It's your comprehensive evaluation framework. This "eval" is a suite of tests and benchmarks that defines what "good" looks like across multiple dimensions (e.g., accuracy, latency, tone, fairness). The team's job is no longer to "ship features"; it's to continuously improve the system's score against that evaluation framework.

• Embrace the Probabilistic Mindset. This is the fundamental mental upgrade required. Stop thinking in binary terms of "right" and "wrong." Start thinking in confidence scores, accuracy thresholds, and acceptable failure rates. Your job is to define the boundaries of acceptable performance and guide the system towards that state.

The Way Forward

These five mistakes are five facets of the same core lesson: product leadership is not a static set of rules. It's a dynamic practice of adapting your mindset to the problem at hand. The playbooks that lead to success in one era can become the anchors that sink you in the next.

My goal in sharing these failures is to help you shortcut the learning process. Avoid these traps, and you'll be ahead of 90% of the field. You'll be free to focus on the real work: building products that matter.

In the last post, we went over the new operating system for product management of AI-native products: The Eval-Driven Development framework. To get more concrete, let’s go through the practical steps for building an evaluation framework for a hypothetical product: an "AI Sales Assistant."

The assistant's primary function is to draft personalized follow-up emails for sales representatives based on data from the company's CRM and transcripts of recent sales calls. The overarching business goal is to increase sales rep efficiency while improving the quality and consistency of customer communications.

Applying the five-step EDD playbook to this product would look like this:

• Objective: The product objective is not merely to "build an email drafter." A measurable objective would be: "Generate a factually accurate, relevant, and on-brand follow-up email draft in under 30 seconds that a sales rep can send with minimal editing."

• Dataset: To create the "golden set," the team would gather 200 real (and fully anonymized) CRM opportunity records, each with corresponding call notes or transcripts. Then, the company's top three performing sales reps would be tasked with hand-writing the "perfect" follow-up email for each of these 200 scenarios. This collection of expert-written emails becomes the "ground truth" against which the AI's outputs will be measured.

• Metrics & Framework: The core of the EDD process is the multi-dimensional evaluation framework. This is not a simple checklist but a detailed scorecard that balances different aspects of performance. The table below illustrates what such a framework could look like.

This framework translates abstract goals into a tangible, measurable system. The dimensions chosen—Functional Correctness, Response Quality, User Trust & Safety, and Business Impact—directly map to the key challenges of building a successful AI product.

Each metric is concrete and tied to a specific measurement method, demonstrating the practical application of the hybrid evaluation approach. For instance, "Tool Call Accuracy" is a deterministic check perfect for a fast, code-based eval. "Relevance" and "Coherence" are subjective and well-suited for a scalable LLM-as-judge. "Tone Alignment" is highly nuanced and best calibrated with expert human evaluation before being automated. Finally, "User Adoption" and "Edit Distance" are business-level metrics tracked via behavioral analytics that measure whether the product is actually delivering value to users. This structured table is more than an example; it is a template for translating the abstract theory of EDD into a practical tool that any product team can adapt and implement.

The Real Moat: Your Evals Are Your Most Defensible Asset

In the rapidly commoditizing landscape of AI, traditional competitive moats are eroding. If the underlying technology and the surface-level features are no longer defensible, where does a sustainable competitive advantage come from?

It is surely not the model or the features. Instead, it is an AI company's proprietary, ever-improving eval framework and the unique "golden dataset" that powers it. This is a fundamental shift in strategic thinking.

Evaluation frameworks are highly defensible for several key reasons:

• They Encode Unique Domain Expertise: A well-constructed eval is the codified expression of a team's deep, nuanced understanding of its specific customers and their definition of "quality."

An evaluation framework built to assess the outputs of a legal tech product is useless for a company building an AI for medical diagnostics. This domain-specific judgment, embedded in the evals, cannot be easily copied or reverse-engineered.

• They Create a Powerful Data Flywheel: A product designed with a probabilistic mindset actively captures user feedback and corrections.

This feedback is used to continuously refine and expand the golden dataset. A better, more diverse dataset leads to more robust evals. More robust evals enable the team to build a better, more reliable model. A better model delivers a superior product experience, which in turn attracts more engaged users, who generate more high-quality feedback data.

This creates a virtuous, self-reinforcing cycle—a data network effect that is powered and accelerated by the evaluation engine.

• They Drive Superior Iteration Speed: A team with a robust, automated evaluation system can test hypotheses, measure improvements, and ship better products far faster than a competitor relying on manual checks, subjective opinions, and gut feel. In the fast-moving world of AI, the speed and quality of iteration is a decisive competitive advantage.

The fundamental mindset shift for product leaders, therefore, goes beyond mere process change. It is a strategic reorientation. The job is no longer to manage a backlog of features to be shipped. The new mandate is to act as the chief architect of an evaluation engine. That engine—the system that embodies the company's unique judgment and powers its learning loop—is what will separate the winners from the losers.

The game is no longer about having the best algorithm today; it is about building the best system for improvement tomorrow. This is the core of the probabilistic mindset, and it is the key to building not just a successful AI product, but a lasting, defensible business.

The Mindset That Got You Here Won't Get You There

When I was at Google Deepmind, I led a team of product managers building the first generation of Gemini experience. I was shocked to find that our decades of experience that made us top-tier product leaders seemed to become our biggest vulnerability.

For years, a seasoned product manager would be rewarded for building solid, thoughtful and predictable product experiences. This deterministic mindset, honed over countless successful launches, is now a liability.

This isn't an exaggeration. The failure rate for enterprise AI projects is shockingly high—some estimates say over 80% of them fail. This isn't a technology problem indeed. It’s a mindset problem, a direct result of applying an old playbook to a new game.

Modern AI is inherently probabilistic. With a given input, you wouldn’t know the exact output. This isn't a bug; it's a core feature. But it breaks every rule in the traditional PM playbook. Leaders demand certainty from systems designed to operate on probability, and products collapse under the weight of this flawed expectation.

The solution is not a new tool. It’s a fundamental upgrade to your own mental operating system. It requires a "Probabilistic Mindset Shift"—the move from thinking in binary right/wrong answers to thinking on a different set of quality measures, likelihoods, confidence scores, and acceptable ranges.

This shift changes the questions you ask. Instead of "Does the feature work?", you must now ask:

• What are the different dimensions and measurements of quality?

• How do you evaluate the non-deterministic outcomes?

• What’re the non-acceptable outputs and how can you prevent them?"

• How do we improve without a single "right" answer?"

Since you can't guarantee perfect correctness, you must design for trust. This isn't a vague goal; it's a concrete design principle with clear strategies like providing explainability, showing confidence scores, and giving users control to override the AI. User trust becomes a core metric you can and should measure.

This new mindset leads to a new operating system called Eval-Driven Development.

This is the playbook for that OS shift, drawn from my in-the-trenches experience leading products at Google DeepMind and coaching the next generation of AI-native product leaders and entrepreneurs.

From Feature-Driven to Eval-Driven

The traditional operating system for product development is Feature-Driven Development (FDD). It operates on a straightforward contract: the Product Requirements Document (PRD) specifies, in detail, the features to be built. The team then executes against those specifications. Success is measured by shipping the defined features with high accuracy, on time and within budget. This model is perfectly suited for the predictable, deterministic world of traditional software.

However, when applied to AI development, the FDD model breaks down completely:

• Performance Cannot Be "Specced": A PM cannot write a PRD that states, "The AI will generate 'good' marketing copy." The definition of "good" is subjective, contextual, and cannot be captured in a static requirements document. The quality of an AI's output exists on a spectrum, not as a binary pass/fail state.

• Outcomes Are Inherently Unpredictable: A team does not know how well a model will perform a novel task until they actually build and test it. And if you test it with the traditional QA methods, you don’t get what you actually look for. The AI development lifecycle is not a linear march toward a known endpoint; it is an uncertain and highly iterative process of research and discovery.

• The Focus Is Wrong: FDD concentrates on what to build—the feature. Effective AI development must focus on how well the system performs a task—the outcome.

The failure of the old model necessitates a new operating system: Eval-Driven Development (EDD). In this paradigm, the central governing artifact of the product process is no longer the PRD. It is the evaluation framework, or "eval" for short. The core idea of EDD is that the product requirement is not a list of features to be built. Instead, the requirement is a set of clearly defined, measurable criteria that constitute a "good" or successful output for a given task.

The development cycle is transformed from a linear Spec -> Build -> Test -> Ship process into a continuous, data-driven loop:

Define Eval Criteria -> Test Against Criteria -> Analyze Failures & Gaps -> Improve -> Repeat

The product manager's role undergoes a transformation accordingly. The primary job is no longer writing feature specs. It is architecting the evaluation framework and system itself. The PM's product judgment, deep understanding of user needs, and definition of quality are encoded directly into the evaluation framework, which then guides the engineering and data science teams' iterative work.

The adoption of EDD represents more than just a change in process; it triggers a fundamental power shift within a product organization. It moves the source of truth and authority from static, narrative-based documents (PRDs) to dynamic, data-driven systems (eval frameworks). In the FDD world, power is concentrated in the "what we should build" decision, and the product manager, as the author of the PRD, is the primary gatekeeper of that decision.

In an EDD world, the eval framework becomes the source of truth. Its creation is an inherently collaborative act, PMs need to build deep partnership with data scientists to define statistically sound metrics (like precision, recall, or F1-score), with engineers to build the automated testing infrastructure, and with domain experts to provide the nuanced, qualitative judgment of what "good" actually looks like in a specific context.

Consequently, the product manager's role evolves to defining the problem space through the architecture of the evals and then managing a portfolio of experimental bets to improve performance within that space. This requires a different set of skills: less feature specification and more statistical literacy, experimental design, and deep technical collaboration.

The EDD Playbook: A Step-by-Step Guide

Moving from theory to practice requires a concrete, actionable playbook. The following five-step process outlines how to implement Eval-Driven Development, transforming it from an abstract concept into a day-to-day operational reality for product teams.

Step 1: Define Your Objective (Not Your Feature)

The process begins by anchoring on the core user problem and the desired business outcome, not on a proposed solution. In traditional development, a team might receive a request to "Build a feature to summarize call transcripts." In EDD, this is reframed as an objective: "Reduce the time our sales reps spend on post-call administrative work by 50% by providing them with accurate and relevant summaries." This objective-first orientation is a cornerstone of modern product management, but it becomes non-negotiable in the world of AI, where the path to the solution is not dictated in advance.

Step 2: Collect Your Dataset (The "Golden Set")

An evaluation framework is useless without high-quality data to test the system against. The next critical step is to assemble a "golden set" of representative examples that will serve as the benchmark for performance. This dataset is thoughtfully curated to reflect the full spectrum of real-world usage. A robust golden set should include:

• Happy Paths: Typical, common scenarios that represent the core use case.

• Edge Cases: Uncommon but plausible inputs that test the boundaries of the system's capabilities.

• Adversarial Cases: Inputs deliberately designed to trick, confuse, or "jailbreak" the system, testing its safety and robustness.

The data for this set can be sourced from various places, including anonymized production logs, examples manually curated by domain experts, or even synthetically generated data designed to cover specific scenarios. The guiding principles for this dataset are diversity and realism.

Step 3: Define Your Eval Metrics (The Multi-Dimensional Scorecard)

This step is the heart of the EDD process. A single, high-level metric like "accuracy" is almost always insufficient and can be misleading. Instead, a multi-dimensional scorecard is needed to capture a holistic view of the AI's performance. This scorecard is typically built using a hybrid of three distinct evaluation methods:

1. Code-based Eval: This method uses automated code to perform objective, deterministic checks. It is fast, cheap, and ideal for verifying things like output formatting ("Does the output contain valid JSON?"), length constraints ("Is the summary under 200 words?"), or the presence of specific keywords.

2. Auto Eval (LLM-as-Judge): This innovative technique uses another powerful LLM (like GPT-4o or Claude 3.5 Sonnet) as an automated "judge" to score more subjective qualities. By giving the judge model a clear rubric, it can assess dimensions like "relevance," "clarity," or "coherence" on a numeric scale. This approach offers a scalable way to evaluate subjective quality but requires careful prompt engineering for the judge model to ensure consistency.

3. Human Eval: This remains the gold standard for evaluating the most nuanced and subjective qualities, such as adherence to a specific brand voice, creativity, factuality, or overall user trust. While it is the most time-consuming and expensive method, it is essential for establishing a "ground truth" and for calibrating the automated LLM-as-judge evals. Given its heavy lifting, plan ahead and manage this process well to drive efficiency and quality.

A comprehensive scorecard will include metrics across several dimensions, such as Factual Accuracy, Relevance, Coherence, Tone, Safety, and Lack of Bias.

Step 4: Run and Compare Evals (The Iteration Engine)

With the objective, dataset, and metrics in place, the team can begin the core iteration loop. This is a scientific process of experimentation, not a gut-feel-driven one. The loop proceeds as follows:

1. Establish a Baseline: Run the initial version of the AI system (the "V0") against the full evaluation suite to establish a baseline score for every metric on the scorecard.

2. Formulate a Hypothesis: Propose a specific, testable change. For example: "I hypothesize that adding a chain-of-thought reasoning step to the prompt will improve the 'Factual Accuracy' score by 10% without significantly degrading the 'Latency' score."

3. Implement and Re-run: Make the single change to the system and re-run the entire evaluation suite.

4. Analyze the Results: Compare the new scorecard to the baseline. Did the 'Factual Accuracy' score improve as predicted? Did the change cause an unexpected regression in another area, like the 'Tone Alignment' score?

This data-driven feedback loop replaces subjective debates with objective evidence, allowing the team to iterate rapidly and make demonstrable progress.

Step 5: Continuously Evaluate (Closing the Loop)

EDD does not end when the product is launched. It is a continuous process.

The production environment must be instrumented to constantly gather feedback and surface new failure modes. User feedback—both explicit (ratings, correction submissions) and implicit (tracking when a user ignores a suggestion)—should be collected and analyzed. The most interesting examples, especially failures, should be fed back into the "golden set," making it a living, growing asset.

This continuous loop of evaluation and refinement is the only way to prevent model performance from drifting or degrading over time and to ensure the product becomes more valuable, not less, after it ships.

In the next post, we are going to dive deeper into a case study to see how this Eval-Driven Development framework manifests itself in practice. We are also going over how you would turn your eval framework to be the most defensible moat for your AI product. Stay tuned.